Software Issues

• GCC compiler (gcc, g++)

  	-O2 -fomit-frame-pointer -funroll-loops -malign-double
  	

• PGI compiler (pgcc, pgCC)

  	-O2 -Mvect -p p6
  	

• BLAS, FFT, etc.

The way to reset the precision of FPU operations is

    #ifdef __linux__
       __setfpucw( (unsigned short)((_FPU_DEFAULT & 0xF0FF) | _FPU_DOUBLE) );
    #endif
    

1. Lightweight MD (MDBNCH)
   This highlights weak optimization capabilites of g77 Fortran compiler

   MACHINE, COMPILER, COMPILATION OPTIONS               TIME    DATE
   --------------------------------------------------------------- ------  -------
   
   DEC AlphaStn 500/500(EV5 500MHz/8MB),f77 4.1 [1] ............... 6.93s  17Jun97
   DEC AlphaStn 500/500(EV5 500MHz/8MB),f77 4.1 [2] ............... 7.17s  17Jun97
   DEC AlphaStn 500/500(EV5 500MHz/8MB),f77 4.1 -fast-speculate all 7.28s  17Jun97
   [1] -fast -speculate all -tune host -feedback x.fb
   [2] -fast -speculate all -tune host
   
   SGI Origin 2000 R10000 (195MHz/4MB SC), f77 7.1 [^] ...........  7.65s+ 11Dec96
   [^] -64 -O3 -mips4 -align64 -r10000 -OPT:Ol=0:IEEE_a=3:ro=3:pad=on
       -TENV:X=3 -LNO:fis=0:pref=0:cs2=4M -GCM:ag:ar -CG:body_ins=0
       -INLINE:=ON:preemp -TARG:pl=ip27
   
   ...
   Dell XPS H266 266MHz Pentium II, Linux, Portland pgf77 1.6 [*]  19.9 s  29Sep97
   Dell XPS H266 266MHz Pentium II, Linux, Portland pgf77 1.4 [^]  20.4 s  21Aug97
   [*] -O2 -Minline -Mvect=smallvect -Munroll -Minfo -tp p6 -pc 64
   [^] -O2 -Mvect=altcode -Munroll -tpp6
   ...
   Dell PentiumPro 200MHz/256K, Win95, Intel F77 2.4, -G6 -Qxi ... 22.6 s  22Oct97
   ...
   PentiumPro 200MHz/256K cache, Linux, g77 0.5.18+gcc 2.7.2 [*] . 29.6 s  25Apr97
   PentiumPro 200MHz/256K cache, Linux, g77 0.5.18+gcc 2.7.2 [^] . 33.0 s  25Apr97
   [*]  as [^], plus -fstrength-reduce -fthread-jumps -mno-ieee-fp
   [^]  -O3 -malign-double -funroll-all-loops -fomit-frame-pointer
        -ffast-math^M
   ...
   Dell XPS,I.PentiumPro200MHz,WindowsNT 3.51,djgpp 2.0+f2c+gcc[^] 31.6 s  10Apr96
   [^] gcc -c -O3 -m486 -ffast-math -fstrength-reduce
           -fthread-jumps -mno-ieee-fp
   ...
   Pentium 586 200MHz, 256k cache, Win95, MS PwrStatn 4.0, -G5 -Ox 46.7 s  08Nov96
   Pentium 586 200MHz, 256k cache, Win95, MS PwrStatn 4.0 ........ 76.6 s  08Nov96
   ...
   PC, PPro 200MHz/256Kb, OS/2 Warp 4.0, g77 0.5.21 [^] .......... 47   s  10Nov97
   [^] -O6 -funroll-all-loops -fstrength-reduce -fthread-jumps
       -fomit-frame-pointer -ffast-math -malign-double -fno-automatic^M
   ...
   Micron Pentium Pro 200MHz, Linux, g77 2.7.2.f.1 [^] ........... 49.5 s  18Jul96
   Micron Pentium Pro 200MHz, Linux, g77 2.7.2.f.1, -O6 .......... 51.6 s  18Jul96
   Micron Pentium Pro 200MHz, Linux, g77 2.7.2.f.1, -O ........... 52.6 s  18Jul96
   Micron Pentium Pro 200MHz, Linux, g77 2.7.2.f.1 ............... 83.1 s  18Jul96
   [^] -O3 -ffast-math -fstrength-reduce -fthread-jumps -mno-ieee-fp  ^M
   ...
   Intel i860 40 MHz (board Microway on a 386 PC) [^] ............ 172 s    13Feb91
   [^] analog of the Paragon node.
   		

   
   
2. Full-featured MD (CHARMM)
   In this benchmark a single PentiumPro is roughly equivalent to 8 Paragon nodes

   Nodes             1      2      4      8      16     32     64    128    256    512
   
   HP/735           2.93   1.49   0.78   0.48
   
   Cray-T3E                1.47   0.73   0.37   0.18   0.094  0.054  0.034  0.025
   SGI-Origin       1.19                 0.17   0.095  0.055
   Cray-T3D                                     0.77   0.40   0.21   0.13   0.086 0.068
   Cray-C90         0.51   0.263  0.14   0.074
   IBM/SP2          1.93   0.98   0.50   0.26   0.14   0.09
   Intel Paragon   14.37                 1.92   0.98   0.52   0.29   0.18   0.14  0.098
   Intel Delta                                                              0.17  0.13
   DEC-alpha        0.96   0.51   0.27   0.14
   TERRA            5.03          1.00   0.505  0.27   0.15
   IBM/SP1          6.55   3.42   1.78   0.96   0.59   0.28   0.22   0.19
   CM-5E                          2.24                 0.32   0.20
   SGI P. Challenge 2.88   1.46   0.78   0.38   0.21
   SGI/R10000       1.61   0.82   0.42   0.22
   Convex CSPP-1000 2.50   1.27   0.69   0.39   0.22
   Intel Gamma     20.72  10.49   5.31   2.66   1.37   0.73   0.42   0.26
   CM-5                           6.10                 0.90   0.63   0.42   0.30
   Cray-J90         2.00   1.038  0.565  0.343
   PentiumPro       2.04   1.05
   HP-J200          2.63   1.39
   
   Cray Y-MP        1.01
   Convex - 3       3.11
   
   HP/C180          1.24
   IBM-560          7.45
   

   
   
3. Quantum mechanics (GAMESS)
   Pentium II is not that slower than SGI machine that was considered high-end not so long ago.

   
   

   Machines  (details)	Clock	C4H4	C4H4	C4H6	Ti2H8	Thymine	Thymine
   		MCSCF	MCSCF	GVB	MP2	RHF	RHF
   		grad Full NR	grad SOSCF	hessian	energy	gradient	gradient, Direct
   SGI 4*R8k   R8000/75MHz, 4 CPUs.	CPU  (0.94)	696  (0.76)	****  (****)	607  (1.06)	1224  (0.98)	1045  (0.97)	****  (****)
   Pentium 2/266   266MHz, Gateway(?) with 128MB memory. Red Hat Linux 4.2. SPECfp95= 7.68	CPU  (0.74)	608  (0.66)	****  (****)	530  (0.93)	745  (0.60)	845  (0.79)	****  (****)
   SG Power Challenge   R10000/194MHz. SPECfp95= 13.80	CPU  (0.53)	512  (0.56)	****  (****)	346  (0.61)	603  (0.48)	499  (0.46)	****  (****)
   DEC 500/500   AXP5/500MHz. SPECfp95= 20.40	CPU  (0.39)	328  (0.36)	****  (****)	263  (0.46)	439  (0.35)	437  (0.41)	****  (****)